註冊並分享邀請連結,可獲得影片播放與邀請獎勵。

Mika Okamoto
@mikahokamoto
Agents @DecagonAI, prev @twosigma AI explainability, safety, and evals @gtcomputing
加入 August 2026
53 正在關注    64 粉絲
We're releasing PACT, a benchmark for rule-following under pressure in enterprise AI assistants. One sentence of pressure raised violations 65% across 23 models. None is reliable enough to run unsupervised. Paper and leaderboard: Data:
顯示更多